AI V&V Notebooks

A collection of Jupyter notebooks demonstrating AI/ML verification and validation (V&V) techniques, focusing on adversarial robustness and model reliability.

Notebooks

Notebook	Description
adversarial_traffic_light.ipynb	Demonstrates adversarial attacks on image classification. Uses Foolbox to craft imperceptible perturbations that fool a pretrained ResNet-18 into misclassifying a traffic light as a tree. Highlights risks for safety-critical systems like autonomous vehicles.
ML_robustness_for_demand_prediction.ipynb	Compares ML model robustness for demand prediction. Trains neural networks and uses PyCaret for AutoML, then applies certified robustness analysis via auto_LiRPA to bound prediction changes under input perturbations. Includes a Gradio demo and W&B experiment tracking.

Setup

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Then launch Jupyter:

jupyter notebook